Achievable closed-loop properties of systems under decentralized ...

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO. 5, MAY 1994

932

Achievable Closed-Loop Properties of Systems Under Decentralized Control: Conditions Involving the Steady-State Gain Peter J. Campo and Manfred Morari, Member, ZEEE

which must be specified in each SISO design is typically much smaller than a full multivariable design. 4) Simplijied Tuning. Individual subsystems can be (manually) tuned and retuned on-line to accommodate the effects of (slowly) changing process conditions. The requirement that the control system be decentralized introduces the pairing problem. The pairing problem is concerned with defining the control system structure, i.e., which of the available plant inputs is to be used to control each of the plant outputs. For a fully noninteracting plant, the choice is obvious, and the benefits of decentralized control discussed above accrue trivially. In any practical problem, there are (to a greater or lesser extent) interactions in the I. INTRODUCTION plant. This implies that even if the control system is deESPITE the closed-loop performance advantages of mul- centralized, subsystems of the closed-loop system are not tivariable controllers, the use of single loop controllers independent of each other. To the extent that the control for multivariable plants is the rule in industrial process control system can be designed to make the closed-loop subsystems applications. In addition to its inherent simplicity, a decentral- independent, the idealized characteristics outlined above can ized control system consisting of independent controller sub- be realized. systems (typically single input-single output (SISO) control When the process interactions are significant, the choice of loops) exhibits several advantages over a fully multivariable a control system structure is far from trivial and has been the design. In the ideal case these advantages include: subject of much research [17], 1241, 1251, [28], [331, 1351. For Flexibility in Operation. A decentralized structure allows an n x n plant there are n! possible SISO loop pairings. For operating personnel to restructure the control system by plants beyond even a modest number of inputs and outputs, a bringing subsystems in and out of service individually. brute force approach (to design controllers for every possible This flexibility allows the system to handle chang- pairing and then select the design which provides the best ing control objectives during different operating condi- closed-loop performance) is impractical. tions (for example start-up, shutdown, temporary process This complexity drives the need for analysis methods to modifications to accommodate maintenance, etc.). determine achievable closed-loop system characteristics as a Failure Tolerance. The flexibility characteristics of the function of control system structure independent of controller decentralized structure also provide advantages in the design. With these tools, pairings which do not admit accase of unanticipated structural changes in the control ceptable closed-loop performance can be discarded before system. In particular, if an actuator or sensor fails, only any controllers are designed. The development of such tools the individual subsystem involved is affected. Only this and their use in the pairing problem is the subject of this subsystem need be taken out of service with no changes paper. An important characteristic of the results developed to other parts of the control system. here is that they only require steady-state knowledge of Simplijied Design. Sequential design of individual SISO the plant. This information is easily obtained from simple subsystems is usually easier than a full multivariable identification experiments or steady-state design models. Often design. In particular, the number of design parameters the pairing question can be answered without the need for detailed dynamic modeling which, in itself, can be expensive Manuscript received March 26,1992; revised May 10,1993. Recommended and time consuming for large scale systems. by Past Associate Editor, J. S. Freudenberg.

Abstract-The question of the existence of decentralized controllers for open-loop stable multivariable systems which provide particular closed-loop properties is investigated. In particular, we study the existence of decentrakd controllers which provide integral action ('Ispe 1 closed-loop performan&) and also demonstrate one or more ofi u n c d d o n a l stability, integrity with respect to actuator and sensor failure, and decentralized unconditional stability. Necessary, sufftaent, and, in some cases, necessary and sufl&ient conditions on the open-loop steady-state gain are derived such that there exists a controllerwhich provides these desired closed-loopcharacteristics.These resultsprovide the basis for a systematic approach to control structure selection for decentralized controller design.

D

P. J. Campo is with General Electric Corporate Research and Development, Schenectady, NY 12301 USA. M. Morari is with the Department of Chemical Engineering, Califomia Institute of Technology, Pasadena, CA 91125 USA. IEEE Log Number 9216624.

'Even if design and performance evaluation could be completed in one second, this process would require loo0 hours to complete for a plant with 10 inputs and 10 outputs.

0018-9286/94$04.00 0 1994 IEEE

CAMPO AND MORARI: SYSTEMS UNDER DECENTRALIZED CONTROL

933

Several authors have reported on work in this area. Grosdidier and Morari [16] took a similar approach to the study of single loop controllers for multivariable plants. They defined the concepts of integral stabilizability and integral controllability which are generalized here. Skogestad and Morari [33] introduce the important notion of decentralized integral controllability (DIC) and give some necessary and some sufficient conditions for a system to be DIC. The work of Morari and coworkers is nicely summarized in [27]. Chiu and Arkun [7] study the problem of failure tolerance in a similar setting. In this paper, we bring these ideas together in a unified way and present a number of novel results which generalize those available to date. 11. PRELIMINARY DEFINITIONS

The general system under study is shown in Fig. 1. We make the following assumptions throughout the paper. The n input n output plant, G ( s ) ,is finite dimensional, linear time invariant, stable, and strictly proper. Type 1 closed-loop behavior (i.e., integral action) is required for all n plant outputs. The controller, K ( s ) , is finite dimensional, linear time invariant, proper, and of the form 1

K ( s ) = -C(S) S

(1)

where C ( s ) is diagonal, stable, and such that C(0) is nonsingular, i.e., K ( s) is fully decentralized. The limitation to open-loop stable plants is not restrictive since we will be interested in studying the behavior of the closed-loop system as subsystem controllers are (arbitrarily) taken out of service; in this setting it is not meaningful to consider open-loop unstable plants. The requirement of integral action is typical in process control applications and allows us to connect closed-loop system properties to the open-loop steady state gain. Given assumptions 2) and 3), the inclusion of an explicit integrator in each channel of K ( s ) as in (1) is without loss of generality (note that C ( 0 ) nonsingular is necessary to have integral action on all outputs). Most of the results to be presented can be extended to block decentralized controllers, i.e., those for which C ( s ) is block diagonal, using concepts such as the block relative gain [24] and block D-stability [l], [22], although this is not pursued here. The requirement that C(s) be stable imposes a certain loss of generality but this is not significant in most process control applications. Because the controller K ( s ) is assumed to be diagonal, the control system input-output pairings are determined by the definition of G(s). To consider all possible input-output pairings we must consider all transfer function matrices, G(s), given by

Fig. 1. The closed-loop system under study.

Given this introduction we are prepared to introduce a classification of closed-loop systems along the lines of the flexibility and on-line tuning properties discussed in Section I. Definition I: The system in Fig. 1 is internally stable if bounded signals injected at any point in the system give rise to bounded signals at all other points in the system. We will use the terminology “ K ( s ) stabilizes G(s),” to mean that the given controller and plant form an internally stable closed-loop system when connected as in Fig. 1. Definition 2: The ^closed-loop system in Fig. 1 is unconditionally stable if K ( s ) = E K ( s ) stabilizes G ( s ) for all E E €1 where A €1 =

{ E = (YII (Y

E

(0, l]}.

(3)

Unconditionally stable closed-loop systems allow a minimum measure of on-line tuning. In particular, the gains of each of the subsystem controllers can be simultaneously varied by a factor in the range (0, 11 and the system will remain stable. Unconditional stability can also be regarded as a measure of closed-loop robustness since for a particular controller, K ( s ) , unconditional stability implies that stability is maintained if the elements of G(s) vary by a scalar factor as well. Definition 3: The Aclosed-loop system in Fig. 1 demonstrates integrity if K ( s ) = E K ( s ) stabilizes G ( s ) for all E E Ello where A

= { E = diag(6i)

I ei

E (0, l}, i = l , . . . , n } .

(4)

This definition of integrity implies that the closed-loop system remains stable as subsystem controllers are arbitrarily brought in and out of service (see, for example, [31]). A related definition of closed-loop integrity [ 151 requires the system to remain stable in the face of arbitrary sensor or actuator failure^.^ Integrity in this sense requires that both K ( s ) and G(s) be stable, which precludes Type 1 closed-loop behavior and is, therefore, too restrictive for our purposes. On the other hand, integrity in the sense of Definition 3 does not imply sensor or actuator failure tolerance unless the failure is recognized and the affected control loop taken out of service. Dejnition 4: The closed-loop system in Fig. 1 is decentralized unconditionally stable if K ( s ) = E K ( s ) stabilizes G(s) for all E E ED where A

ED = { E = diag(6i)

I ~i

E [0, 11,

i = l , . . . , n }.

(5)

where P is a n x n permutation matrix.*

A closed-loop system which is decentralized unconditionally stable allows the gains of each controller subsystem to be modified independently by a factor in the range [0, 11. Note that the gains can be made zero which corresponds to taking a subsystem out of service. Thus decentralized unconditional stability implies closed-loop integrity.

Recall that each column of a permutation matrix consists of 1 one and n - 1 zeros. Furthermore, permutation matrices are orthogonal so that P-’ = P T .

With K ( s ) diagonal, integrity with respect to sensor failures is equivalent to integrity with respect to actuator failures.

G(s) = P G ( s )

(2)

934


While these definitions are useful for classifying closedloop systems, they are of no direct use in the pairing problem because they require the specification of a controller, K ( s ) , a priori. To overcome this limitation, we define a parallel classification of open-loop systems. In particular, we classify a given open-loop system, G ( s ) ,according to whether or not there exists a controller of the form (1) which together with G(s) demonstrates a particular closed-loop property. Formally we define:

A theorem-proof approach is adopted for rigor, with numerous remarks following each result explaining its significance and offering illuminating examples. All new results are proven in detail; proofs of known results are provided only when they are not readily available elsewhere or when the proof adds clarity to the following exposition. Notation is introduced as required and summarized in Appendix A.

ti(S).

to be definitive-more sophisticated closed loop analysis tools, such as the real multivariable gain margin or the structured singular value, could be used to study the closed-loop problem. Our purpose here, however, is only to state those results which are required to for the more interesting open-loop classifications in Section IV.

DeJnition 6: The plant G ( s ) is integral controllable (IC) if there exists a controller, K ( s ) ,of the form (1) such that the system in Fig. 1 is unconditionally stable. Dejinition 7: The plant G ( s ) is integral controllable with integrity (ICI) if there exists a controller, K ( s ) , of the form (1) such that the system in Fig. 1 demonstrates integrity. DeJnition 8: The plant G ( s ) is decentralized integral controllable @IC) if there exists a controller, K ( s ) , of the form (1) such that the system in Fig. 1 is decentralized unconditionally stable. Since only existence of a suitable controller is required in Definitions 5-8, the dependence on K ( s ) inherent in Definitions 1-4 is removed. In Section IV we develop analysis methods which allow us to classify a given G ( s ) according to Definitions 5-8 without having to explicitly construct K ( s ) . Since the classifications defined by Definitions 5-8 depend on the choice of controller structure through the definition of G ( s ) ,these analysis methods allow us to study the suitability of various potential control structures without having to complete detailed controller designs. In particular we can determine if given closed-loop flexibility and on-line tuning characteristics are achievable for a particular plant and controller structure independent of controller design. Morari [26], Grosdidier and Morari [16], and Morari and Zafiriou [27] present more limited definitions of integral stabilizability and integral controllability. Their definitions involve properties of a closed-loop system with a given K ( s ) and correspond to the definitions of internal stability and unconditional stability presented here. Our motivation in generalizing these definitions is to make them independent of K ( s ) and consistent with the definition of decentralized integral controllability originally introduced by Skogestad and Morari [33]. The remainder of the paper is devoted to developing conditions on G ( s ) which allow us to classify it according to Definitions 5-8. These results take the form of necessary, sufficient, or, where possible, necessary and sufficient conditions for membership in a particular class. Because a number of conflicting definitions and results have appeared in the literature, we include results for each of IS, IC, ICI, and DIC. Our main focus, however, is on the richer system properties IC1 and DIC. The primary reasons for this are the engineering significance and relative absence of existing results for these classes. To make the presentation clearer and more self contained, known results are interspersed with new results as appropriate.

HI. CLOSED-LOOP SYSTEM PROPERTIES

A. Internal Stability

The following well-known result provides a necessary and sufficient condition for internal stability of the system in Fig. 1. Fact 1: Given that G ( s ) is stable, the closed-loop system of Fig. 1 is internally stable if and only if Q(s) 2 K ( s ) [ I G ( s ) K ( s ) ] - lis stable. This result is a special case of the Youla parameterization of all stabilizing controllers (see for example [14]). For a discussion and proof with the notation used here see [27].

+

B. Unconditional Stability Using a [ - ]to denote the spectrum of a matrix argument and

c+- (0) to denote the closed right-half plane less the origin, the following holds from Theorem 3 of [26]. Fact 2: The system of Fig. 1 is unconditionally stable only if a[G(O)C(O)]c E+ - (0). Fact 3: Given C ( s )diagonal and stable, there exists y > 0 sufficiently small such that the system of Fig. 1 with K ( s ) = :C(s) is unconditionally stable if a[G(O)C(O)] C C+. Except in the case that G(O)C(O)has purely imaginary eigenvalues, Facts 2 and 3 completely characterize unconditionally stable closed-loop systems. Example 4 of Grosdidier and Morari [161 demonstrates that nothing can be said about unconditional stability from G(O)C(O) alone when its spectrum contains purely imaginary values. C. Integrity

To study the situation where parts of the controller can be taken out of service, we need the notions of principal submatrices and principal subsystems. The index set M consisting of k tuples of integers in the range 1,.. . , n is defined by A

M =

{(il,.-.,ik)

11

5 il < iz... < i k 5 n}.

(6)

Given an n x n matrix A, each m = {il,. . . ,ik} E M defines a k x k principal submatrix of A, denoted A,, made up of the rows and columns of A indexed by m. For a given m E M, the notation, dim (m), refers to the number of elements in m


and indicates the dimension of the principal submatrix, A,. Given a system represented by the transfer function matrix G ( s ) ,its principal subsystems are represented by the transfer function matrices G , (s), the principal submatrices of G(s). Given this notation the following is a direct consequence of Definition 3. Fact 4: The closed-loop system of Fig. 1 demonstrates integrity if and only if K , (s) stabilizes G , (s) for all m E M.

935

B. Integral Controllability

The following necessary condition follows from Fact 2. Theorem 2: G ( s )is integral controllable only if there exists a diagonal matrix, X , such that a[G(O)X]c C+ - (0). Proof: We establish the result by contradiction. Assume that no such X exists but that G ( s ) is IC. This implies that there exists a K ( s ) of the form (1) with C(0) nonsingular such that K ( s ) unconditionally stabilizes G ( s ) .By Fact 2 this implies that a[G(O)C(O)] c - {0}, but this contradicts the D.Decentralized Unconditional Stability assumption. U The following simple result follows from Definitions 2-4. In a similar fashion we can use Fact 3 to develop a sufficient Fact 5: The closed-loop system of Fig. 1 is decentralized condition on G(0) such that G ( s ) is IC. unconditionally stable if and only if for all diagonal D > 0 and Theorem 3: If there exists a diagonal matrix, X , such that all m E M, (1/11D112)[DK(s)],unconditionally stabilizes o[G(O)X]c C+, then G ( s ) is IC. Gm(4.4 Proof: By Fact 3, under the given conditions there exists Trivially, decentralized unconditional stability implies that a y > 0 such that the controller K ( s ) = ( y / s ) X uncondiK ( s ) stabilizes G ( s ) with integrity and that K m ( s )unconditionally stabilizes G ( s ) .The existence of such a K ( s ) implies tionally stabilizes G,(s). This follows directly from Fact 5 in that G ( s ) is IC by definition. 0 the special case D = I. Remarks: 1) The condition of Theorem 3 implies not only that G(s) is Iv. OPEN-LOOPSYSTEM PROPERTIES IC, but also that there exists a “pure integral” controller of the form, K ( s ) = ( y / s ) X , with X a constant We now turn our attention to the characterization of openmatrix, which demonstrates that G ( s ) is IC i.e., which loop systems. These results allow us to answer the question, unconditionally stabilizes G(s). “For a given plant and variable pairing defined by G ( s ) ,does 2) The reverse direction of Theorem 3 is known not to hold. there exist a controller of the form (1) which together with (Example 4 of [16] provides a counterexample.) G ( s ) forms a closed-loop system with property X?’ where property X is any one or more of internal stability, un- Any condition on G(0) which implies that there exists an X conditional stability, integrity, or decentralized unconditional satisfying the condition of Theorem 3 implies that G(s) is stability. Since the answer depends upon the variable pairing IC. The least restrictive such result known is provided by the chosen, these results provide us with quantitative measures for following lemma. evaluating the suitability of potential control system structures Lemma I : Given any matrix, A, if there exists a permutawithout having to develop controller designs. tion matrix, P , such that the leading principal submatrices of PAPT are nonsingular, then there exists a diagonal matrix, X , such that a [ A X ]c C+. Furthermore, if the determinants A. Integral Stabilizability of the leading principal submatrices of PAPT are positive, X A complete characterization of the IS property is provided may be chosen to be positive definite. by the following theorem involving an easily verified condiProof: That there exists such an X was proven origition. nally by Fisher and Fuller [13]. An alternative proof due to Theorem 1: G ( s )is integral stabilizable if and only if G(0) Ballantine [3] demonstrates that when the determinants of the is nonsingular. leading principal submatrices are positive, X may be chosen Proof: This result is basically contained in Theorem 3 0 to be positive definite. of Davison [8]. Necessity is easily demonstrated using Fact 1 Using Lemma 1, the following corollary follows directly above. Sufficiency follows from the observation that the augmented plant, ( l / s ) G ( s ) can , have no unstable decentralized from Theorem 3. Corollary I: If there exists a permutation matrix, P , such 0 fixed modes. that the leading principal submatrices of PG(0)PT are nonRemarks: singular, then G ( s ) is IC. 1) The condition that G(0)be nonsingular, and hence the IS Proof: By Lemma 1, under the given conditions there property, is pairing independent. In fact, the necessity of exists a diagonal X such that o[G(O)X]c C+. By Theorem a nonsingular steady state gain matrix can be established 0 3 this implies that G ( s ) is IC. for any controller of the form (l), even if C ( s ) is a full Remarks: transfer function matrix, i.e., even if the controller is 1) Corollary 1 may be equivalently restated as: If there not decentralized. exists a sequence of nonsingular principal submatrices, 2) The condition that G(0) be nonsingular is equivalent G m l ( 0 ) ,GmZ(0),...,G,,,(O), with G m k ( 0 of ) order IC to the requirement that the plant have no transmission and mi c mj V i < j, then G ( s ) is IC. zeros at the origin. 2) Theorems 2.2 and 2.3 of Guardabassi et al., [18] provide 4Throughout the paper we use 11. I I p to denote the matrix norm induced by an equivalent result with a more involved proof. the Holder p-norm on 72”.

c+


936

Except in the case that there exists a diagonal matrix X by5 which provides u[G(O)X]C - {0}, but no such X which S A 2 diag {sign ( a l l ) , sign (a22),. ,sign (a,,)} (9) provides o[G(O)X]c C+, Theorems 2 and 3 uniquely classify any given system as either IC or not IC. and for any real number, 2, While it is known that the reverse direction of Theorem 3 1 ifxLO does not hold, it is believed that the reverse direction of Thesign (x)= (10) -1 i f x < O ’ orem 2 does. In the absence of any proof or counterexample we state this as the following. Finally we introduce the relative gain array (originally Conjecture I : If there exists a diagonal matrix, X , such that used tool in the pairing defined by Bristol [51) a a[G(O)X]c - {0}, then G ( s ) is IC. selection problem. We close this section with the following result which relates Given a nonsingular matrix A, the relative Definition systems which are IC to those which are IS. gain array (RGA) is defined to be the matrix Theorem 4: There exists a permutation matrix, P , such that P G ( s ) is IC if and only if G ( s ) is IS. A ( A ) A @ [A-’IT (1 1) Proof: Given that P G ( s ) is IC, we have from Theorem 2 that PG(0) must be nonsingular. Since every permutation where @ denotes element-by-element multiplication of matrimatrix is nonsingular, this implies that G ( 0 ) is nonsingular, ces (often called the Hadamard or Schur product). The diagonal elements of A[G(O)]are commonly referred and therefore G ( s ) is IS. Given that G ( s ) is IS, Theorem 1 implies that G ( 0 ) is to as the relative gains of the system G ( s ) . The following well-known properties of the RGA are easily nonsingular. Using Gaussian elimination we can derive from G(0) nonsingular matrices P, L, and U, where P is a proven (see, for example, [16]). Lemma 2: For any nonsingular matrix A, the following permutation matrix, L a lower triangular matrix, and U an upper triangular matrix such that P G ( 0 ) = LU. Elementary hold: calculation then provides that 1) &,(A) = (ui,A’J/det(A)), where A,, is the ( 2 , j)th cofactor of A.6 2) PlA(A)P2 = A(PlAP2) for any permutation matrices Pi and P2. 3) A ( A ) = h(X1AX2) for any nonsingualr diagonal matrices X I and X , . where mf. denotes the leading principal submatrix of P G ( 0 ) With these preliminaries we can prove the following result of dimension T . Since L and U are nonsingular it must hold that Zii # 0 and uii # 0 for all i E { l , . . - , n } . It which leads directly to a necessary condition for ICI. Lemma 3: For any matrix A the following are equivalent: follows then that all leading principal submatrices of PG(0) are nonsingular. By Corollary 1, this implies that the system 1) A+ E P. P G ( s ) is IC. 0 2) There exists a diagonal matrix, X , such that A X E ’P. Remark: 3 ) A(Am) exists and its diagonal elements are positive for all m E M. The existence of a permutation such that P G ( s ) is IC is Proofi We show that 1)+2)+3)+1). equivalent to saying that there exists an input-output pairing By definition, A+ = ASA, so that A+ E P implies that such that G ( s )is IC. This means that if G ( 0 )is nonsingular, we can always find at least one pairing for which there exists there exists an X , namely X = SA, such that AX E P. Thus 1) implies 2). Next we show that 2) implies 3). AX E P a controller which unconditionally stabilizes G (s). implies that A, is nonsingular for all m E M so that existence of A(A,) is established. By Lemma 2, we have C. Integral Controllability with Integrity

c+

~

c+

Before developing conditions for ICI, we introduce some additional definitions. Definition 9: The matrix A is a P-matrix if det A , > 0 for all m E M. In the obvious way we define the set, P , consisting of all P-matrices.

P 4 { A I detA, > 0 for all m

EM}

(8)

so that we may write, A E P to mean that A is a P-matrix. We also introduce the notation, A+, to denote the matrix derived from A by postscaling to make its diagonal elements nonnegative. Definition IO: Given the matrix A, we denote by A+ the matrix, A+ = ASA, where the “sign matrix,” SA, is defined

&,(A,) = & , ( A X ) , for all m E M aii~ii(AX): for all m E M . det (AX),

(12)

We see then that Aii(A,) consists of three factors, aiixii, (AX):, and l/det ( A X ) , , each of which is positive since AX E P. Thus 2) implies 3). Finally we show that 3) implies 1). We have

A;;(A,) > 0 for all m E M

(13)

Aii(A2) > 0 for all m E M

(14)

which implies

5Notice that by definition S A is nonsingular and Si’ = S A for any A. 6For A of dimension 1 we take A l l = 1 by convention.

931


where T = dim (m). Since d,(s) have no roots in E+ only if

so that

uA(AA)aZ det (A&) Since a;

>0

for all m E M .

(15)

(A$&)ii > 0 for all m E M . det (AA)

QSUb(S) = K,(S)[I -k G,(s)K,(s)]-l

(17)

stable for all m E M . With G ( s ) stable this implies that

I - Gm(S)Qsub(S) = [I+ Gm(s)K,(s)]-'

(18)

must be stable for all m E M . This implies in turn that the polynomial equation det [I has no roots in N,(s) where

+ G,(s)K,(s)]

=0

c+.Substituting G,(s)K,(s) d,(s)

= ais;

(19) = (l/sd,(s))

+ . . . + u1s + 1

is the (stable) pole polynomial of G,(s)C,(s), is a polynomial matrix, we have det [sd,(s)l+ N,(s)] = 0 srd& (s)

(20) and N,(s)

(21)

(22)

c+- (0). Expanding the determinant in (22)

s'd&(s)+...+det[N,(O)]

Equation (16) allows us to prove that AS E P by induction. We first establish that det (A$&,)> 0 for dim ( m l )= 1. Then assuming that det (A$&J > 0 for dim (m,) = T , we show that (16) implies that det(A$&p+l) > 0 for dim(m,+l) = T + 1. This completes the induction and establishes that det (ALJ > 0 for all dim (m,) = T E (1, . . . ,n}, or equivalently that A+ E P. Consider first the case that T = 1. The principal submatrices of A+ of dimension 1 are simply its diagonal elements, U:. Since these are positive by construction of A+, we have det(AAl) > 0 for all m E M such that dim(m1) = 1. We now assume that for some fixed T > 1, det (ALJ > 0 holds for all m, E M . This assumption implies that the diagonal cofactors of AAp+,, denoted are positive since these cofactors are simply the determinants of principal submatrices of of dimension T , i.e., (A&p+l)ii = det (ALP) > 0. With (ALp+l)ii > 0 for all i, (16) implies that det(ALr+,) > 0 for all m,+1 E M . The induction is then established, and it must hold that det (A;?) > 0 for all m, E M such that T E { 1,. . . ,n}. This is equivalent to the condition A: E P for all m E M . Thus we have established that 3) implies l), and the proof is complete. 0 We are now in a position to state a necessary condition for ICI. 7'heoremS: G ( s ) is IC1 only if G(0) satisfies any (and hence all) of the conditions of Lemma 3. Proof.. Using Fact 4, G ( s ) IC1 implies that there exists a controller, K ( s ) = (l/s)C(s),such that K,(s) stabilizes G,(s) for all m E M . This implies by Theorem 1 that G,(O) must be nonsingular for all m E M . Furthermore by Fact 1 we must have

z+,(21) will

det [sd,(s)I+ N,(s)] = 0 has no roots in we have

> 0 by construction, we must have

is nonzero on

=O.

(23)

c+

If this polynomial is to have no roots in - {0} then by Routh-Hunvitz, all coefficients in (23) must be nonnegative. This implies that det [Nm(0)]= det [G,(O)C,(O)] 2 0.

(24)

We have det Cm(0) # 0 by assumption and we have previously shown that det G,(O) # 0 so that we require det [G,(O)C,(O)]

>0

(25)

for all m E M . This implies that there exists a diagonal matrix, X = C(O),such that G ( 0 ) X E P and the result is proven. 0 Remarks: Theorems 3.3 and 3.4 of [7] can be reworked to form a necessary condition for IC1 equivalent to condition 3) of Lemma 3. The reverse direction of Theorem 5 holds for 3 x 3 and smaller systems (we state and prove this in Section VI below). It is conjectured that the reverse direction does not hold in general although no proof of counterexample has been demonstrated. A somewhat surprising consequence of Theorem 5 is the following corollary. Corollary 2: G(s) is integral controllable with integrity only if G(s) is integral controllable. Proof: G ( s ) IC1 implies that G+(O) E P which implies that all of the principal submatrices of G(0) are nonsingular. By Corollary 1 this implies that G ( s ) is IC. 0 This result says that if there exists a controller which together with G(s) forms a closed-loop system with integrity that there must also exist a controller which together with G ( s ) forms an unconditionally stable closed loop. This is somewhat unexpected since the sets Ello and €1 are unrelated. An obvious corollary to Theorem 5 is the condition implicit in Theorem 6 of [26] that the relative gains of G ( s ) must be positive for integral controllability with integrity. Corollary 3: G ( s ) is integral controllable with integrity only if Aii[G(O)] > 0, i = 1,.. . ,n. While it is true that positive relative gains are necessary for controllability with integrity, this condition is not sufficient. In fact, as the following example shows, positive relative gains are not even sufficient for the weaker condition that G ( s ) be IC. Example 1: Let G ( s ) be given such that r l 1 1 11

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO. 5 , MAY 1994

938

positive relative gains, Xis E .178)- By construction* Principal submatrix Of G(O) is singular, so that G+(o) and G(s) is IC'* it be using the Routh-HurWitZ criteria that there exists no diagonal x such that .[G(O)X] c E+ - (0). Thus by Theorem 2, G(s) is not IC. 5) Chang and Davison [6] state as Theorem 1 of their pawr that G+(o) E is necessary and sufficient for the existence of a controller, K ( s ) = ( l / s ) X , with X a nonsingular constant diagonal matrix such that the following hold: which

has

5*oo9

a) b)

.1787

K ( s ) unconditionally stabilizes G(s). Every SISO principal subsystem of K ( s ) stabilizes the corresponding principal subsystem of G(s).

That G+(o) E P is not necessary is easily shown with the following example. Example 2: Let 1 1 1

G(s) =

['

s + l 1 0 21]

(27)

v. DECENTRALIZEDINTEGRAL CONTROLLABILITY preliminary to the study of systems which are DIC, we introduce the notion of D-stability, which was originally defined by Arrow and McManus [2]. The definition currently in use in the literature is slightly less restrictive than Arrow and M ~ M and~ is ~ ~ -here. Definition 12: The real matrix A is D-stable if and only if C+ for all diagonal matrices > 0. .(AD) For convenience we define the set of all D-stable matrices DL

e { A I .(AD)

> 0)

(28)

e { A I a ( A D ) c e+V diagonal D > 0).

(29)

c C+

V diagonal D

and the related set

DL

With a slight abuse of notation, we will write A E - (0) to indicate that A is nonsingular in addition to being in the set i.e., A E E - (0) if and only if a(AD) C - (0) V diagonal D > 0. With these ideas understood, we Can State the following necessary condition for the system G(s) to be DIC. - (0) for Theorem 7: G(s) is DIC only if GL(0) E all m E M. Proof: By Fact 5, G(s) is DIC if and only if there exists a K ( s ) = (l/s)C(s) with C(s) diagonal and C(0) nonsingular such that ( 1/11D112)[DK(~)]m unconditionally stabilizes Gm(s) for all m E M and for all diagonal D > 0. Using Fact 2 this implies that

m,

e+

G(0) is clearly not a P-matrix since G,(O) is singular for m = (1, 2). However, a[G(O)] = {.1154 f .5897i, 2.769) c C+, so that by Fact 3 there exists an -y > 0 such that K ( s ) = (T/s) unconditionally m stabilizes G( s). That K , (5) also stabilizes G, (s) when E M , and > This imp1ies that dim (m) = 1 is also clear. Thus the required condition, for G+(O) E P, cannot be necessary. (31) 4 [ G ( o > D c ( o ) l m )c - (0) We note that the condition G+(O) E P is sufficient for properties a) and b) to hold. In fact it can be shown using for all m E M , and D > 0. Writing G(O) and c ( 0 ) as Lemma 1 and Fact 3 that the weaker condition there exists a (32) G(0) = G+(0)S$o) = G+(O)SG(~) permutation matrix, P , such that all leading principal submatrices of PG+(0)PT have positive determinants, is sufficient (33) C ( 0 )= c+(o)s;;o) = SC(O)C+(O) for the stated properties to hold. A sufficient condition for G(s) to be integral controllable and substituting into (31) provides with integrity is given by the following theorem. . ~ [ ~ + ~ ~ ~ ~ ~ ( o ) c ~ c (-o(0)) ~ + (34) ~ ~ ~ ~ l m ~ Theorem 6: If there exists a diagonal matrix X such that a[G(O)X], c C+ for all m E M, then G(s) is ICI. for all m E M, and D > 0. For (31) to hold with D = I proofi By Theorem under the given conditions for and dim (m) = 1, it must hold that g a 2 ( 0 ) ~ a ( O>) 0 for all every m E M there exists a ~ ( m )> 0 sufficiently small i E (1,.. . ,n ) . It follows then that sG(o)sc(o) = 1. n u s if such that the K m ( s ) = ( T ( m ) / s > X m uncondi(34) is to hold, we must haye g([G+(O)B,],) c E+ - (0) tionally GVZ(s) for E M . By taking y* = for all E M , where D = C+(O)D is an minmcMT(m)9 we define the K ( s ) = (?*/'Ix positive definite diagonal matrix. This implies by definition which, by construction, demonstrates that G(s) is ICI. 0 - {o) for all 0 that Gk(0) E E M. Remarks: Remarks: ') The condition Of Theorem imp1ies not Only that G(s) 1) All known necessary conditions for G(s) to be DIC folis ICI, but also that there exists a pure integral controller, low from Theorem 7 as special cases. Specific examples K ( s ) = (r*/s)X, with X a constant matrix, which include the following. together with G(s) forms a closed-loop system with Corollary 4 [27]: G(s) is DIC only if both of the integrity. following hold: 2) In light of Remark 1) it is conjectured that the reverse direction of Theorem 6 does not hold although no a) a[G+(O)D] c E+ for all diagonal D 2 0. specific counterexample has been demonstrated. b) The relative gains of G(s) are positive.

e+

e+

39

cmm AND MORARI: SYSTEMS UNDER DECENTRALIZEDCONTROL

939

Proofi Condition a) follows trivially from Theorem 7. Condition b) is established as follows. By Theorem 7, G(s) is DIC implies that Gk(0) E - (0) for all m E M . It follows immediately that G+(O) E P. By Lemma 3 this implies that the relative gains of G+(s) must be positive which, in turn, implies that the relative gains of G ( s ) must be positive. U 2) The condition of Theorem 7 (along with a) of Corollary 4) is not finitely verifiable so that it is not useful as a practical test for G ( s ) to be DIC. The best computable necessary condition obtainable from Theorem 7 is apparently that the spectrum of G+(O)D and all of its - { 0) for a specific principal submatrices must lie in D 7 0 (or a finite number of such Ds). For example, with D = I Theorem 7 becomes the following. - Corollary 5: G ( s ) is DIC only if a([G+(O)],} c C+ - ( 0 ) for all m E M . Note that this implies, but is not implied, by G+(0) E P, a necessary condition for G ( s ) to be integral controllable with integrity. 3) It is conjectured that the reverse direction of Theorem 7 does not hold although no proof or counterexample is known. The concept of D-stability can also be used to develop a very general sufficient condition for DIC. Zkorem8: If G L ( 0 ) E D L for all m E M , then G ( s ) is DIC. Pro08 It follows from the given condition that a{ [G(O) sG(O)D],} c C+ for a11 m E M and all diagonal D > 0. This implies that there exists a nonsingular matrix X , namely X = S G ( ~such ) , that a[G(O)DX],C C+ for all m E M and all D > 0. It follows from Theorem 3 (Remark 1)) that for all m E M and for all diagonal D > 0 there exists a y > 0 such that K,(s) = ( r / s ) X , unconditionally stabilizes [G(s)D],. By Fact 4 and Definition 3 this implies that for all diagonal D > 0 and for all E E Ello, K ( s ) = ( r / s ) E X stabilizes G ( s ) D This, in turn, implies that for all diagonal D > 0 and for all E E E l p , k ( s ) = ( y / s ) D E X stabilizes G ( s ) . Since the set of all E = D E where D is positive definite and E E Ello contains ED, this implies that G ( s ) is DIC. 0 Remarks: 1) The condition of Theorem 8 implies not only that G(s) is DIC, but also that there exists a pure integral controller of the form, K ( s ) = ( y / s ) X ,with X a constant matrix, which demonstrates that G(s) is DIC. 2) In light of Remark l), it is not surprising that the condition of Theorem 8 is known to not be necessary for DIC. We demonstrate this with the following example. Example 3: For the plant

eigenvalues of G(0) are a[G(O)]= {k&z, 3) so that G(0) is not D-stable. Note that although G(0)has eigenvalues on the imagfor all D > 0 and for inary axis, a[G(O)D],c all m E M so that the necessary conditions of Theorem 7 are fulfilled. As in the tests for IC, the gap between the necessary conditions of Theorem 7 and the sufficient conditions of Theorem 8 occurs along the imaginary axis-when G+(0) and its principal submatrices are in DL{O) but not in DL. 3) In Theorems 1 and 2 of [30], Nwokah et al. claim that the condition of Theorem 8, G,(O) E D L for all m E A4,is necessary both for K ( s ) = (l/s)I with G ( s )to provide decentralized unconditional stability, and for G ( s )to be DIC. Example 3 above demonstrates that this condition cannot be necessary for either decentralized unconditional stability or DIC since K ( s ) = (l/s)I and G ( s ) constitute a decentralized unconditionally stable closed loop while G(0) is not D-stable. A complete characterization of matrices for which A , E D L for all m E A4 is provided by Hartfiel [19]. Lemma 4: Given the n x n matrix A , with a(A,) C C+ for all m E M , the following are equivalent: 1) A, E D L for all m E M. 2 ) A is iqterior to the set DL. 3 ) p ( D , A) > 0 for all diagonal D > 0 where

'This involves a straightforward but tedious application of the Routh-Hurwitz stability criteria.

8By convention we take the determinant of an empty matrix (zero rows or columns) to be one.

c+

c+-{O}

p ( D , A ) = det [I + ( D A ) 2 ] -

(36)

c

a /O E M dim ( - ) = d i m ( 4 ) = 0 , 1.

det [A%OAP,aDa,aDP,PI .n

(37)

where Aa, p denotes the (not necessarily principal) submatrix of A with rows indexed by a and columns indexed by p.' In the vocabulary of [l], the condition that A lie in the topological interior of D L is equivalent to the statement that A is strongly D-stable. Using this result we can restate Theorem 8 as Corollary 6. Corollary 6: If for all m E M , a[G&(O)] c C+ and G$(O) satisfies any (and hence all) of the conditions of Lemma 4, then G ( s ) is DIC.

While Lemma 4 provides necessary and sufficient conditions on G+(O) such that the condition of Theorem 8 is satisfied, the condition of the resulting Corollary 6 is no more easily verified than that of the theorem. On the other hand any condition which implies that GL(0) E D L for all m E M trivially generates another (possibly conservative) Corollary (35) to Theorem 8. For an overview of available results of this kind, the interested reader is referred to the excellent survey by Johnson [20] and references therein. it can be shown that the controller, K ( s ) = (l/s)I, Among the known conditions which imply that all principal results in a decentralized unconditionally stable closed- submatrices of a given matrix are D-stable, we will focus on loop system.' Thus, by construction, G ( s ) is DIC. The the following.


940

Lemma 5: The following conditions are equivalent and imply that A , E D L for all m E M. 1) There exists a diagonal F > 0 such that AF+FAT > 0. 2) (I A)-l exists and there exists a diagonal F > 0 such that (IF(I- A ) ( I A)-lF-l112 < 1. Proof: That 1) implies A E DL is due to Arrow and McManus [2]. The extension to A , E D L for all m E M can be found in [20]. See also [4] for a discussion. It remains to show that 1) is equivalent to 2). We first note that 1) implies that u ( A ) c C+ by Lyaponov’s theorem so that (I A ) - l must exist. Given existence of (I A)-’, it follows that:

+

+

+

+

+ A)-lF-l1I2 < 1 * X,,[(I + F A F - l ) - T ( I - FAF-l)T .( I - FAF-l)(I + FAF-l)-l] < 1 H ( I + FAF-l)-T(I - FAF-l)T .( I - FAF-l)(I + FAF-l)-l < I

IIF(I - A ) ( I

(38) (39) (40)

( I - FAF-l)T(I - FAF-l) < ( I + FAF-l)T(I FAF-l) (41) H - ( F A F - l ) T - FAF-l < ( F A F - l ) T + FAF-l

U

+

(42) (43) (44)

e F A F - l + (FAF-l)T > 0 U A F - ~+ F - ~ > A o~

4) The most obvious consequence of Corollary 7 is that any system which has an (adjusted) steady-state gain matrix, G+ (0), with positive definite symmetric component is DIC? Although significantly more restrictive than Theorem 8, or even Corollary 7, this previously unknown, and obviously simple, condition has proven useful in several practically motivated examples.” A number of simple sufficient conditions for DIC follow from Corollary 7 and are stated in terms of the “error system,” L ( s ) , defined (whenever it exists) by L(8) = [ G ( s )- G ( S ) ] E - ’ ( S )

(46)

where c(s)= diag { g ; ; ( s ) } . The norm of L ( s ) is a measure of the relative error incurred by approximating G ( s ) by its diagonal elements. Using Corollary 7 we can easily prove the following. Corollary 8: If L(0) exists and A

min

F>O, diagonal

IIFL(O)F-’~[~

O, diagonal Remarks: 1) Khalil [21] provides an algorithm which given any It remains then only to show that condition (50) implies that G+(O)is guaranteed to find an F satisfying condition 1) G ( s ) is DIC. By the triangle inequality we have of Lemma 5 if one exists. A practical drawback is that (51) F>O,min diagonal l(-FL(0)F-l - F-1LT(0)F/12< 2. the algorithm need not terminate in finite time. A more attractive numerical approach is to solve This implies that min IIF[I - G+(O)][I G+(0)]-1F-1112 (45) p+FL(O)F-l - F-lLT(O)F] < 2 (52) F>O, diagonal

IIFL(o)F-~~~,

+

to satisfy condition 2) of Lemma 5. Much research has focused on the solution of this optimization problem in connection with the evaluation of an upper bound on the structured singular value [ 113, and several algorithms are available for its solution (see, e.g., [12, 321). These algorithms make use of the fact that (45) can be reformulated as a convex optimization problem for which descent directions are provided by the solution of an eigenvalue problem. 2) A significant advantage of Corollary 7 is that since it involves a single test on G+(O),no further examination of the principal submatrices of G+(O)is required. This is particularly important for large systems since the number of principal submatrices grows exponentially with n. 3) A disadvantage of Corollary 7 compared to Corollary 6 is that it is more conservative. Hartfiel [19] provides a 3 x 3 example with A , E D L for all m E A4 for which no F exists to satisfy 1) of Lemma 5.

where p 2 ( - ) denotes the measure of its matrix argument defined by the induced 2-norm. (See [9, 101 for a definition of p p ( . ) and a discussion of its properties). Using the definition of L(O), this implies

,U~[--FG(O)E-’(O)F-~ - F - ’ ~ - ’ ( 0 ) G T ( O ) F< ] 0 (53) which in turn implies that

G(0)E-1(O)F-2+ F-2E-1(0)GT(O)> 0.

(54)

Since G(O)G-l(O) = [G(0)G-l(O)]+this implies by Corollary 7 that the system, G ( ~ > E( 0-)is ~ DIC. This in turn implies that G ( s ) is DIC, and the proof is complete. 0 9Recall that the symmetric component of the real matrix A, denoted S ( A ) , is given by S ( A )= (1/2)(A A T ) . “See, for example, those in [23]-[25], [29], [33]. Example 5 of [33] is trivially shown to be DIC using this condition, while in the referenced paper an intensive numerical search is only able to provide that DIC is “extremely likely.”

+

I


94 1

1

FDLTl and stable

Theorem 9: For G(s) of dimension 2, the following are equivalent.

'll[G(0)1 > 0. G+(0) E P. G(s) is DIC. G(s) is ICI. Proofi We prove that 1) j 2 ) j 3 ) j 4) + 1). That 1) 3 2) is evident from Lemma 2 which provides

1) 2) 3) 4)

Fig. 2. Classification of general systems.

It follows from Corollary 8 that if G(0) is (generalized) diagonally dominant, Le,. p((L(O)() < 1, then G(s) is DIC." This is not surprising since diagonally dominant systems are "weakly coupled" in a precise mathematical sense. As we argued heuristically in the introduction, we expect weakly coupled systems to admit decentralized controllers with attractvie closed-loop properties. What is more interesting is that the system G(s) need not be weakly coupled (diagonally dominant) to be DIC. Indeed, as the next example shows, Corollary 8 can be arbitrarily conservative relative to Corollary 7 (which is itself more conservative than Theorem 8). Example 4: Consider any G(s) such that (55) L

where a E

J

R is given. Since for any a G'(0)

In the 2 x 2 case, G+(O) E P implies that G z ( 0 ) E D L for all m E M [20], so that by Theorem 8, 2) implies 3). That 3) implies 4) is trivial from the definitions of IC1 and DIC. That I7 4) implies 1) follows from Theorem 5. In [34], Yu and Fan claim to provide a necessary and sufficient condition for DIC in the 3 x 3 case. Unfortunately the proof contains a flaw. We restate their result here in corrected form as Parts 1) and 2 ) of Theorem 10.'' Part 3 ) of Theorem 10 is a new result. Theorem 10: For G(s) of dimension 3, 1) If G ( s ) is DIC, then G+(O) E P and E:='=, xi/2[G(0)] 2 1. 2) If G+(0) E P and X,1,!'[G(O)] > 1, then G(s) is DIC. 3 ) G(s) is IC1 if and only if G+(0) E P. Before stating the proof of Theorem 10, we note that according to Theorem 2 of [34], condition 2 ) is necessary as well as sufficient for G(s) to be DIC. That this condition cannot be necessary is established by Example 3 for which G+(O) E P and E:=, X;!'[G(O)] = 1 although G(s) is DIC. A careful reworking of the arguments in the proof presented in [34], however, provides the following result. Lemma 6: Let A be a 3 x 3 matrix with A+ E P, then the following hold: ,

I

E - (0) for all m E M implies ':/'(A) 2 1. 2) x ~ / ~ (>AI )implies A A E D L for all m E M . Proofi We first note that for dim(m) 5 2, A+ E P - (0) as well) imlies that AA E D L (and hence AA E [20]. It remains to show that in the 3 x 3 case

1) 1)AA

+ [G+(O)IT= I > 0

(56)

G ( s ) is DIC. For p E (1, 2, m}, however,

which can be made arbitrarily large by the choice of a. We conclude this section on the classification of open-loop systems by summarizing all of the proven results in the Venn diagram of Fig. 2. Some nontrivial regions of the diagram are not currently known to be nonempty. For example, there is no for all m E M known G(s) which satisfies GA(0) E m-(0} yet is not DIC. Nonetheless, the indicated set inclusions are known to hold and the diagram facilitates the visualization of the known results.

~ ( A + Dc) C+

- (0) 3

for all diagonal D

> 0 + XXf!'(A) 2

and

i=l

for all diagonal D VI. SPECIALIZED RESULTS FOR 2

X

2

AND

3

X

3 SYSTEMS

A number of simplifications of the results in Section V are available for low dimensional systems. We collect these here in the following theorems. "This result also follows from the development in [27].

1 (59)

i=l

> 0. (60)

The desired result is obtained by following Yu and Fan, and applying the Routh-Hurwitz criteria to determine the location 0 of the spectrum of A+D for arbitrary D . '2Unfortunately the corrected version no longer provides a necessary and sufficient condition.


942

Among the screening tools for variable pairing obtainable from the results of this paper, the most practically relevant are: 1) The necessary condition for DIC of Corollary 5. 2) The necessary condition for DIC of Theorem 5 based on Condition 3) of Lemma 3. 3) The sufficient condition for DIC of Corollary 7 based on Condition 2) of Lemma 5 . Condition 1) is apparently the tightest known computable condition necessary for DIC. Unfortunately it requires the evaluation of the spectrum of all 2" - 1 principal submatrices of G+(O) for each of the n! possible pairing choices. This is prohibitive except when n is relatively small. Condition 2) is relatively weak, while necessary for DIC it is not even sufficient for IC. On the other hand, property 2) of Lemma 2 provides that by evaluating the relative gain array for the 2" - ( n+ 1) principal submatrices of order 2, . . . ,n13 we can check all of the n! possible pairings; i.e., as rows of G ( s ) are interchanged to study alternative pairings, the only change in the RGA is an interchange of the corresponding rows. The principal advantage of Condition 3) is that it only involves a test on G+(O) and not its principal submatrices.

C(5)is FDLTI and stable

C(s) is DIC

Fig. 3. Classification of 2 x 2 systems.

C(s) is FDLTI and stable *

II

Fig. 4. Classification of 3 x 3 systems. We will continue with the proof of Theorem 10. Proof [of Theorem IO]: Using Lemma 6, parts 1) and 2) of Theorem 10 follow trivially from Theorems 7 and 8, respectively. Part 3) remains to be shown. Given that G ( s ) is ICI, it holds that G+(O) E P by Theorem 5. This implies (by Lemma 1) that there exists a diagonal F > 0 such that u[G+(O)F]c C+. Furthermore, G+(O) E P and F > 0 diagonal imply that G+(O)F E P This implies that a[G+(O)F], c C+ for all m E M. (Any 1 x 1 principal submatrix of G+(O)F is positive by the definition of G+(O) and the fact that F > 0. Any 2 x 2 principal submatrix must have its spectrum in the open right-half plane since both its trace and its determinant are positive). Thus there exists a nonsingular matrix, X = S G ( ~ ) F such , that a[G(O)X], c C+ for all m E M. By Theorem 6 this implies that G ( s ) is ICI.

0 As in the general n x n case these results are summarized in a Venn diagrams shown in Figs. 3 and 4. VII. CONCLUSION Practically motivated classifications of a closed-loop and open-loop systems have been defined. The question of the existence of a decentralized controller which provides a closed loop with desired characteristics is studied in detail. The results here generalize and relate all known conditions, necessary or sufficient, for IS, IC, ICI, and DIC. The important connection between D-stability and DIC is demonstrated, and several examples are provided to illustrate this connection. In addition to the independent theoretical interest of these results, we have laid the foundation for the development of a practical solution to the decentralized control variable pairing problem. Based on steady state gain information we can (essentially) determine if there exists a controller with integral action which will provide a closed-loop system with desired flexibility characteristics, any one or more of unconditional stability, integrity, or decentralized unconditional stability.

APPENDIXA NOTATION

X,,,(A)

The eigenvalue of A which is largest in magnitude. A(A) The Relative Gain Array of the matrix A. The measure of the matrix A. p(A) The spectral radius of the matrix A. p(A) a(A) The (eigenvalue) spectrum of the matrix A. a.. The ijth element of the matrix A. %?. A%? The ijth cofactor of the matrix A . The matrix A scaled to make its diagonal A+ elements positive. Am The principal submatrix of A consisting of rows and columns of A indexed by m. c+ The open right-half plane. The closed right-half plane. c+ - (0) The closed right-half plane less the origin. DL The set of all D-stable matrices. mr An index set of dimension r. P The set of all P matrices. SA The diagonal sign matrix derived from A. The symmetric component of the matrix A. S(A) The matrix formed by taking the absolute values IAI of the elements of A. The matrix norm of A induced by the Holder llAllp p-norm on 2".

c+

ACKNOWLEDGMENT This work was performed while the first author was an Alexander von Humboldt Research Fellow at the Technical University of Munich and would not have been possible without the financial support of the Alexander von Humboldt I3The n RGAs of order 1 are trivially positive.

CAMPO AND MORARI: SYSTEMS UNDER DECENTRALIZEDCONTROL

StiftUng. The authors wish to thank Richard D. Braatz for helpful criticisms and suggestions regarding this work.

943

[25] G. Mijares, J. D. Cole, N. W. Naugle, H. A. Preisig, and C. D. Holland, “A new criterion for the pairing of control and manipulated variables,” AIChE J... vol. 32. ~~.DD. 1439-1449.. 1986. [26] M. Morari, “Robust stability of systems with integral control,” IEEE Trans. Automat. Contr., vol. AC-30, pp. 574-577, June 1985. [27] M. Morari and E. Zafiriou, Robust Process Control. Englewood Cliffs, NJ: Prentice-Hall, 1989. [28] C. N. Nett and H. A. Spang, “Control structure analysis and synthesis: A missing link in the evolution of modem control theories,” in Proc. Amer. Contr. Con$, Minneapolis, MN, June 1987. [29] A. Niederlinski, “A heuristic approach to the design of linear multivariable interacting control systems,” Automatica, vol. 7, pp. 69 1-701, 1971. [30] 0. D. I. Nwokah, A. E. Frazbo, and D. K. Le, “A note on decentralized integral controllability,” Int. J. Contr., vol. 34, pp. 485-494, 1993. [31] D. H. Owens, Feedback and Multivariable Systems. London: Peter Peregrinus Ltd., 1978. [32] A. K. Packard, “What’s New With p : Structured Uncertainty in Multivariable Control,” Ph.D. dissertation, University of California, Berkeley, 1988. [33] S. Skogestad and M. Moarai, “Variable selection for decentralized control,” presented at AIChE Annual Meeting, Washington, Nov. 1988. [34] C. C. Yu and M. K. Fan, “Decentralized integral controllability and D-stability,” Chem. Eng. Sci., vol. 45, no. 11, pp. 3299-3309, 1990. [35] C. C. Yu and W. L. Luyben, “Design of multiloop SISO controllers in multivariable processes,” Indus. Eng. Chem. Process Design and Devel., vol. 15, pp. 498-503, 1986. ~~~

REFERENCES E. H. Abed, “Strong D-stability,” Syst. Contr. Lett., vol. 7, pp. 207-212, 1986. K. J. Arrow and M. McManus, “A note on dynamic stability,” Econometn‘ca, vol. 26, pp. 448-454, 1958. C. S. Ballantine, “Stabilization by a diagonal matrix,” in Proc. Amer. Math. Soc., 1970, vol. 25, pp. 728-734, G. P. Barker, A. Berman, and R. J. Plemmons, “Positive diagonal solutions to the Lyapunov equations,” Linear and Multilinear Algebra, vol. 5 , pp. 249-256, 1976. E. H. Bristol, “On a new measure of interaction for multivariable process control,” IEEE Trans. Automat. Contr., vol. AC-11, pp. 133-134, Jan. 1966. T. N. Chang and E. J. Davison, “Steady-state interaction indices for decentralized unknown systems,” in Proc. 1986 Cont Decision Contr., 1986, pp. 881-887. M. S. Chiu and Y. Arkun, “Decentralized control structure selection based on integrity considerations,” Indust. Eng. Chem. Research, vol. 29, pp. 369-373, 1990. E. J. Davison, “The robust decentralized control of a general servomechanism problem,” IEEE Trans. Automat. Contr., vol. AC-21, no. I, pp. 14-24, 1976. C. A. Desoer and H. Haneda, “The measure of a matrix as a tool to analyze computer algorithms for circuit analysis,” IEEE Trans. Circ. Theory, vol. 19, pp. 480-486, 1972. C. A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output Properties. New York Academic, 1975. I. C. Doyle, “Analysis of feedback systems with structured uncertainties,” in IEE Proceedings Part D, vol. 129, 1982, pp. 242-250. M. K. H. Fan and A. L. Tits, “Characterization and efficient computation of the structured singular value,” IEEE Trans. Automat. Contr., vol. 31, pp. 734-743, 1986. M. E. Fischer and A. T. Fuller, “On the stabilization of matrices and the convergence of linear iterative processes,” in Proc. Cambridge Philosop. Soc., vol. 54, pp. 417-425, 1958. B. A. Francis, A Course in H, Control Theory. Berlin: SpringerVerlag, 1987. M. Fujita, and E. Shimemura, “Integrity against arbitrary feedback loop failure in linear multivariable control systems,” Automatica, vol. 26, no. 6, pp, 765-772, 1988. P. Grosdidier and M. Morari, “Closed-loop properties from steady-state gain information,” Indust. Eng. Chem. Fund., vol. 24, no. 2, pp. 221-235, May 1985. P. Grosdidier and M. Morari, “Interaction measures for systems under decentralized control, Automatica, vol. 22, no. 3, pp. 309-319, May 1986. G. Guardabassi and A. Locatelli, “On the initialization problem in the parameter optimization of structurally constrained industrial regulators,” Large Scale Syst., vol. 3, pp. 267-277, 1982. D. J. Hartfiel, “Concerning the interior of the D-stable matrices,” Linear Algebra Appl., vol. 30, pp. 201-207, 1980. C. R. Johnson, “Sufficient conditions for D-stability,” J. Econ. Theory, vol. 9, pp. 5 3 4 2 , 1974. H. K. Khalil, “On the existence of positive diagonal P such that P A .4TP < 0,” IEEE Trans. Automat. Conrr., vol. AC-27, pp. 181-184, 1982. H. K. Khalil and P. V. Kokotovic, “D-stability and multiparameter singular perturbations,” SIAM J. Contr. Optim., vol. 17, pp. 56-65, 1979. D. K. Le, 0. D. I. Nwokah, and A. E. Frazho, “Multivariable decentralized integral controllability,” Int. J. Contr., vol. 54, no. 2, pp. 481-496, 1991. V. Manousiouthakis, R. Savage, and Y. Arkun, “Synthesis of decentralized control structures using the concept of block relative gain,” AIChE J., vol. 32, pp. 991-1003, June 1986.

+

~

~

L.

~

~~

Peter Campo received the B.S. degree in chemical engineering from Rice University, Houston, TX, in 1984, and the Ph.D. degree in chemical engineering from the California Institute of Technology, Pasadena, in 1989. In 1989 and 1990 he spent 14 months at the Technical University of Munich as an Alexander von Humboldt Fellow pursuing research interests in decentralized control and in the design and analysis of controller implementations providing Antiwindup and Bumpless transfer operation in the face of input saturation nonlinearities. Since 1990 be has been a Control Systems Engineer in the Control System and Electronic Technologies Laboratory of the General Electric Corporate Research and Development Center, Schenectady, New York. Dr. Campo’s current research interests include applications of linear and nonlinear control theory and development of machine vision systems for feedback control of manufacturing processes.

Manfred Morari (M’87) obtained the diploma from ETH Zurich and the Ph.D. from the University of Minnesota, both in chemical engineering. Dr. Morari’s interests are in the areas of process control and design. He is the McCollum-Corcoran Professor and Executive Officer for Control and Dynamical Systems at the California Institute of Technology, Pasadena. He has held appointments with the University of Wisconsin, Exxon R&E and IC1 and has consulted internationally for a number of major corporations. In recognition of his research contributions, he received numerous awards, among them the Donald P. Eckman Award of the Automatic Control Council. He was recently elected to the National Academy of Engineering.